Like a ‘Car Check-Engine Light’: How PG&E Uses Data Analytics, Machine-Learning Algorithm to Predict Electric Maintenance Needs

By Ari Vanrenen

A new machine-learning algorithm is helping PG&E detect early stages of wear and tear on its electric equipment so the utility can take preventive action and reduce risk on its system.

“Essentially, we’re trying to create a car check-engine light. If you see it, you can act before something bad happens to your car. Our goal is to maximize the lifetime of the assets as well as detect any anomalies before they take place and cause outages or other unintended consequences,” said Ana Maria Ñungo a technical product manager on PG&E’s Enterprise Data Management team.

PG&E’s Michael Thibault examining a transformer identified by the algorithm.

Through an Electric Program Investment Charge (EPIC) project — a utility research and development program in California — PG&E is analyzing a diversity of existing data sources such as SmartMeters, Geographic Information Systems (GIS), maintenance and incident reports, as well as meteorological information, looking for patterns. The team has built and fine-tuned a machine learning-based algorithm on top of this data that can recognize the early stages of deterioration that may cause equipment failure.

PG&E successfully completed testing the algorithm in July 2021 and has begun using it more widely.

One initiative within the project is focused on PG&E’s 900,000 radial electric distribution transformers with the goal of trying to predict equipment failure within 30 days. Radial transformers do not have a backup system, which means power may take longer to restore in the case of an outage.

Once the algorithm makes a prediction, engineers review the findings and then field crews go to the site to check the transformer. As of October 2021, over 200 engineering reviews have been completed, with nearly 70% confirming anomalies with the transformer. While no algorithm is perfect, this 70% rate is the type of capability that can revolutionize the utility industry.

“Our crews have begun to investigate transformers in the field based on the model’s prediction and will repair or replace equipment as necessary. It’s exciting that we’re in the field demonstrating how the model can help us reduce wildfire risk and improve reliability for our customers,” said Ñungo.

The PG&E team (L to R): Raef Thabet, Danny Mendoza, Ana Maria Ñungo, Peter Elisher, Shane Buck and Michael Thibault. (Not pictured: Eric Schoenman, Devon Yates, Sabrin Mohamed, Maryam Variani, Will McFaul IV, Aayushi Gupta and JP Dolphin.)

While this process continues, PG&E will continue to make the model smarter and better, and work to transition the algorithm to an operational data product.

“The investments that PG&E is making into our data and analytics capabilities are making it possible for us to develop and deploy these industry-leading products,” said technical lead Devon Yates, a data scientist in PG&E’s Asset Knowledge Management department.

The project is currently being operated by PG&E’s Asset Health and Performance Center within Electric Operations. “I’m pleased to have helped bring the project and vision alive, making its way from concept to production,” said Eric Schoenman, PG&E’s project business lead.

When fully implemented, this project will be another step in PG&E’s journey to a more data-driven approach to understanding the health of its distribution assets and mitigate risk.

To learn more about PG&E’s EPIC program and its various projects, visit www.pge.com/epic.

Email Currents at Currents@pge.com.